A basic model 'Centaur' that predicts human responses in psychological experiments has appeared, based on more than 10 million human choices obtained from 160 studies



An international research team from Germany and the United States has announced ' Centaur ,' a platform model that predicts and simulates human behavior in any experiment. Developed based on Meta's open source large-scale language model '

Llama 3.1 70B ,' Centaur was trained on a dataset containing 10 million human choices from 160 psychology experiments, and is said to be generalizable to experiments in completely new domains, even when the story or structure of the experiment is modified.

A foundation model to predict and capture human cognition | Nature
https://www.nature.com/articles/s41586-025-09215-4



Excited to see our Centaur project out in @nature.com.
TL;DR: Centaur is a computational model that predicts and simulates human behavior for any experiment described in natural language.

[image or embed]

— Marcel Binz ( @marcelbinz.bsky.social ) July 3, 2025 0:34



Centaur
https://marcelbinz.github.io/centaur

Scientists Use AI to Mimic the Mind, Warts and All
https://www.nytimes.com/2025/07/02/science/ai-psychology-mind.html

Modern AI is already capable of many things that were once only possible for humans: for example, it can defeat chess and Go champions, drive a car, predict the three-dimensional structure of proteins, and produce natural, human-like conversations and sentences.

However, at the time of writing, many of the AIs are specialized in only certain areas and are very different from humans. A chess champion can drive a car to the venue by himself, but an AI that is good at chess cannot drive a car. Also, although an AI chatbot can have a smooth conversation with a human, when playing chess, it seems that it can make very basic and strange mistakes, such as moving the piece incorrectly.

Despite these shortcomings, some scientists hope that AI can help us understand the human mind. In a new paper published in the journal Nature, cognitive neuroscientist Marcel Binz of the Helmholtz Munich Medical Research Center in Germany and his team present Centaur, a foundational model that can predict and simulate human behavior in any experiment that can be expressed in natural language.



Centaur is based on Meta's Llama 3.1 70B and is trained using the Psych-101 dataset, which consists of the results of 160 psychology experiments involving a total of more than 60,000 subjects. The experiments in Psych-101 include a variety of experiments, such as 'piloting a spaceship to play a treasure hunting game,' 'memorizing a list of words,' and 'playing two slot machines with different payouts to get as many rewards as possible,' and contain more than 10 million choices made by human subjects.

The team trained the Llama 3.1 70B to act as the subject of each experiment, rewarding it when it responded like a human. 'We basically taught the robot to mimic the choices that human subjects would make,' Binz said.

After training, the researchers tested how well Centaur worked, and it accurately predicted how subjects would respond to experiments that were not included in the dataset. They also modified a game in which subjects piloted a spaceship to search for treasure, but instead piloted a flying carpet to search for treasure. Just like the human subjects, Centaur was able to apply the strategies it had developed in the spaceship to the flying carpet.

Even when asked logical reasoning questions that were not included in the dataset, Centaur tended to answer questions that humans answered correctly, but did poorly on questions that humans found difficult. 'There's a lot of generalization going on here,' Binz said.



'Centaur is really impressive,' said Russ Poldrack, a cognitive scientist at Stanford University who was not involved in Centaur's development. 'It's the first model that can do any kind of task exactly like a human can.'

'Ultimately, we want to understand the human mind as a whole and figure out how all of this is connected,' Binz said. The team is working to increase its database of psychology experiments five-fold and plans to further train Centaur.

Centaur can be downloaded from the Hugging Face page below.

marcelbinz/Llama-3.1-Centaur-70B-adapter · Hugging Face
https://huggingface.co/marcelbinz/Llama-3.1-Centaur-70B-adapter



in Software,   Web Service,   Science, Posted by log1h_ik